High-Coverage 3D Scanning through Online Structured Light Calibration

2014 22nd International Conference on Pattern Recognition
High-Coverage 3D Scanning through
Online Structured Light Calibration
Andrea Albarelli, Luca Cosmo, Filippo Bergamasco and Andrea Torsello
Dipartimento di Scienze Ambientali, Informatica e Statistica
Universit´a Ca’ Foscari Venezia
via Torino 155, 30172 Mestre(VE), Italy
[email protected], [email protected], [email protected], [email protected]
Abstract—Many 3D scanning techniques rely on two or
more well calibrated imaging cameras and a structured light
source. Within these setups the light source does not need any
calibration. In fact the shape of the target surface can be inferred
by the cameras geometry alone, while the structured light is
only exploited to establish stereo correspondences. Unfortunately,
this approach requires each reconstructed point to exhibit an
unobstructed line of sight from three independent points of views.
This requirement limits the amount of scene points that can be
effectively captured with each shot. To overcome this restriction,
several systems that combine a single camera with a calibrated
projector have been proposed. However, this type of calibration
is more complex to be performed and its accuracy is hindered by
both the indirect measures involved and the lower precision of
projector optics. In this paper we propose an online calibration
method for structured light sources that computes the projector
parameters concurrently with regular scanning shots. This results
in an easier and seamless process that can be applied directly to
most current scanning systems without modification. Moreover,
we attain high accuracy by adopting an unconstrained imaging
model that is able to handle well even less accurate optics. The
improved surface coverage and the quality of the measurements
are thoroughly assessed in the experimental section.
I.
Fig. 1. The incomplete coverage problem that affects many structured light
systems. See the text for details. (image best viewed in color)
I NTRODUCTION
Structured light 3D scanners have recently become very
popular due to their constantly improving accuracy and affordability. Many of such devices are based on a pair of
cameras for point triangulation and on a pattern projection
to fix correspondences between observed points. This kind of
configuration requires a rather strict condition to hold for a
point to be triangulated. Namely, it must be illuminated by the
projector and seen by both cameras. The consequences of this
requirement are illustrated in Fig. 1. In the shown scenario, the
relative position of the projector and the camera pair allows to
reconstruct only the portion of the surface represented with the
thick blue line. In fact the remaining part, depicted with the
thin red line, cannot be measured since self-occlusion happens
for at least one of the three required lines of sight. By converse,
if the system was able to work with just one camera the
captured surface would have been much more extended. With
this latter scenario, which implies projector calibration, all the
non-dashed surface in Fig. 1 could have been recovered.
This higher coverage of the imaged surface, combined with
the implicit lower costs, asserts the potential usefulness of
a camera-projector system. Of course those benefits can be
exploited only through an accurate projector calibration which,
aptly, is a very well covered topic within literature. Many
1051-4651/14 $31.00 © 2014 IEEE
DOI 10.1109/ICPR.2014.699
approaches involve the projection of some special pattern
over a known planar target. The underlying idea is to allow
the camera to localize the target, while using the pattern to
map individual projector pixels. This is the case, for instance,
with the approach recently proposed by Chien et al. [1],
where a virtual checkerboard is projected onto a real one
and an iterative correction loop is performed on the projector
parameters until the camera detect a perfect overlap between
the virtual and physical target. Chen and Xi [2] adopt a target
made up of a grid of regular dots, which are easy do detect
and located accurately. The projector rays are located by using
a standard Gray-coded pattern sequence to assign a projector
coordinate to each image pixel. Very similar methods are also
proposed (among many others) by Huang et al. [3] (using dots
and Gray coding), Moreno and Taubin [4] (using checkerboard
and Gray coding), and Audet and Okutomi [5] (using physical
augmented reality tags together with projected tags).
While the combination of physical and projected target
is extensively adopted to calibrate projector-camera systems,
several alternative approaches have been explored. A specially
crafted light sensor, made with optical fibers, is used by Lee
et al. [6]. The individual sensors embedded in the target are
used to discover its location by locally decoding a Gray-coded
4080
binary pattern projected over them. This way there is no need
for the camera to estimate the pose of the target plane. Kimura
et al. [7] drops the need for a special reference object, although
this advantage is obtained by assuming that the camera has
been previously calibrated. The process requires to project onto
a plane of unknown pose and uses projective geometry constraints between camera and projector to optimally calibrate the
latter. A planar featureless target is also used in the more recent
method proposed by Yousef et al. [8]. Here two calibrated
cameras are required and passive stereo is performed onto a
projected checkerboard in order to establish correspondences.
Differently, Okatani and Deguchi [9] demonstrate how to
obtain the homographies that relate a plane, a projector and
a camera when just the intrinsic parameters are known.
Finally, some methods provide for auto-calibration. That is,
the projector calibration happens during the scanning process
itself. Most of these approaches are adapted from classical
uncalibrated stereo techniques [10]–[12] where the second
camera is replaced by the projector. A drawback of these
methods is that the actual scale of the scene cannot be
inferred, however Furukawa and Kawasaki [13] suggested how
to augment the system with a laser pointer to recover the
correct Euclidean reconstruction. Yamazaki et al. [14] propose
an online method that exploits the dense correspondences
produced by structured light to compute radial fundamental
matrix that is then decomposed into intrinsic and extrinsic
parameters for both camera and projector. This way they can
account for some lens distortion, however they do not address
the scale ambiguity.
With this paper we introduce a projector calibration technique that does not require to adopt special devices or targets.
Neither an explicit calibration procedure is needed. In fact, our
approach is an online method that can be performed directly
during the normal system usage. Differently from other online
methods it is able to automatically recover the scene scale and
to deal even with severely distorted lens. However this comes
at the price of requiring two calibrated cameras to be paired
with the projector. While this could seem to be a limitation,
it should be noted that our approach is designed for a very
specific, yet critical, scenario. Namely, our goal is to augment
the coverage of standard 3D scanners that are designed to use
a pair of stereo cameras cupled with non-calibrated projector,
which are very common among commercial products. To
validate our method, we will perform specific experiments to
show that it is indeed able to significantly increase the number
of acquired points. This happens without sacrificing quality
and without any modification for off-the-shelf systems.
II.
H IGH -C OVERAGE 3D S CANNING
The main idea of our approach is to exploit the 3D points
triangulated by two calibrated cameras to get insight about
the projector geometry. Basically, this happens by collecting
among subsequent shots the coordinates 3D points that reproject exactly over the same projector pixel and then using a
simple least square fitting to fix the parameters of the projector
ray associated to that pixel. In order to easily perform this step
and to cope well with commercial quality projector optics, we
adopted the general unconstrained camera model. We already
studied a method for effectively calibrating such model [15]
and now we are extending it to deal with this new application.
A. Unconstrained Camera Model
With the term unconstrained camera model we mean a
completely free imaging model where each pixel (i.e. imaging
sensor for cameras or light emitter for projectors) is associated
to an independent ray. More formally, the ray associated with
camera pixel i can be written as ri = (di , pi ), where di , pi ∈
IR3 represent direction and position of the ray respectively.
These vectors satisfy ||di || = 1, (normalized direction) and
dTi pi = 0 (orthogonal position vector). Any point x in the
ray ri satisfies the parametric equation x = di t + pi for some
t ∈ R.
This kind of model comprises literally millions of free
parameters, thus it is very hard to calibrate with standard
target-based approaches. In [15] we propose a practical method
to calibrate it. This is done by taking several shots of a
computer monitor showing both a vertical and horizontal phase
shift pattern sequence [16]. These is exactly the same type of
phase shift coding that is used during the scanning process.
The monitor is placed in a total of s position, each one
characterized by a pose Θs = (Rs , ts ) where Rs and ts are
the rotation and translation of the monitor reference system
with respect to the camera world. Once the pattern sequences
are decoded, for each pose of the target monitor Θs , we
are able to assign an observed code Co to each camera
pixel i that was inside the reprojection area of the monitor
surface. After estimating the poses Θs (see [15] for details),
it is also possible to compute the expected code for each
pixel Ce(ri |Θs ) = Puv (di tint + pi ) where Puv denotes the
projection onto the (u, v) planar coordinates of the monitor
nT (t −p )
reference frame, and tint = s nTsdi i is the intersecting
s
parameter for the equation of ray ri , i.e., the value such that
di tint + pi lies on the monitor plane.
Under these premises, the best estimate for each ray
ri = (di , pi ) is the one thatminimizes the sum of the
−1 T
squared Mahalanobis lengths
of the residuals
k εΣk ε
s
ε = Ce − Co, where Σi is the (conditional) error covariance
matrix under the given ray-pose combination:
s
s
(Σsi )−1 = I + (cos2 φsi − 1)r i (r i )T
(1)
φsi
is the angle between the target and the monitor. Note
where
that, with the pose parameters Θs at hand, these observed 2D
coordinates can be transformed into 3D points in the camera
coordinate frame. We can divide the residual ε = Ce−Co into
the orthogonal vectors ε = Ce − Co and ε⊥ = Co − Co,
where ε is parallel to r . Clearly, since ε⊥ is orthogonal to
the plane spanned by d and n, the point in r closest to Co is
the one closest to Co . Further, let h be this point, we have
||h − Co||2 = ||h − Co ||2 + ||ε⊥ ||2 .
(2)
It is easy to show that, ||h − Co || = cos φ||ε ||, where
φ is the angle between d and n. Hence, the squared distance
between r and Co equals
d2 (r, Co) = cos2 φ||ε ||2 + ||ε⊥ ||2 = εT Σ−1 ε ,
(3)
thus the generalized least squares formulation with respect
to the target coordinates corresponds to the standard linear
least squares with respect to the 3D points associated with
each ray. The linear least squares problem is then solved by a
4081
¯ + wt, where x
¯ = n1 ni=1 xi
ray with parametric equation x
is the barycenter of the observed 3D points, and w is the
eigenvector of their covariance matrix corresponding to the
smallest eigenvalue.
B. Online Projector Calibration
Our goal is to calibrate the rays of the projector as if
it was a camera that always recaptures the exact pattern
that is projected. This assumption, which is similar to many
other projector calibration methods, allows to implicitly know
exactly the observed codes Co. By contrast, the expected code
Ce must be obtained from the scene. Most approaches found
in literature solve this problem by projecting some pattern on a
known target and using the features of the target as references.
In our case, since we have two calibrated cameras available, we
can obtain the 3D reference points by collecting them during
the scanning process. The triangulation of each point of the
scene happens by finding the same code (computed from the
sequence of patterns [16]) on both cameras. This will produce
3D points that are associated to codes observed by camera
pixels, that in general do not correspond to the expected code
Co for any projector ray. For this reason, we cannot directly
use the 3D points belonging to the acquired surface, rather
we must produce additional points corresponding exactly to
the (virtually) observed codes. This can be done easily by
interpolating the camera rays whose codes encompass the
observed code Co with weights inversely proportional to the
distance from Co of their respective measured codes. To solve
the ray interpolation problem for the unconstrained camera
model, we generalize bi-linear interpolation to the manifold
of 3D lines. Under our parametrization, in fact, a line is
represented as a point in IR6 . However, the normal direction
and orthogonal position constraints force the lines to lay in a
4-dimensional manifold. We can generalize (weighted) means
over a manifold through the notion of Fr´echet means: a point
x residing in manifold M is the average of points xi ∈ M
with weights wi if it solves
wi d2M (x, xi )
(4)
argmin
x∈M
i
where dM is the geodesic distance over M. A similar approach
has previously been applied to the interpolation of rotations
and rigid transformation [17], [18] and, similarly to those
approaches, it can be shown that the interpolation is invariant
to the frame of reference and interpolates the rays through a
minimal path with constant linear and angular velocity. For
performance reasons, just like in [18], we approximate the
Fr´echet mean by taking a linear average in IR6 followed by a
projection onto the manifold. These newly obtained virtual rays
can finally be used to triangulate the expected point Ce corresponding to Co and use it to perform the same optimization
described in section II-A. Note, however, that equation 3 holds
only for locally planar surfaces. Generally speaking, this is not
guaranteed if the points Ce are obtained from the scanning of
random objects. Still, since both man-made and natural objects
usually exhibits several low frequency areas, it is reasonable to
guess that at least a sizeable portion of the obtained expected
points will be accurate enough. Moreover, the scanned object
will likely move during different shots, enhancing the coverage
of the projector frustum and eventually adding redundancy.
Finally, even if some Ce could suffer from bad estimation, in
the next section we suggest an apt method to filter outliers.
C. Outliers Filtering
After estimating the projector rays, we can assess their
quality by means of their fitting residual. Specifically, we can
set a badness threshold that can be used to deem as unreliable
rays that obtain a bigger residual. Such rays can be removed,
in which case they will be simply unavailable for triangulation
(note that there is no need to estimate all of the projector rays).
Otherwise, it is indeed possible to still recover them by filtering
inaccurate Ce points that could possibly be the cause of the
bad estimation. To do this, we crate a tentative ray candidate by
applying equation 4 with its available neighbours. Afterwards,
this candidate is used to gather associated Ce that are near a
given threshold to it, which in turn can be used to obtain a new
estimate for the original ray. The rationale of this method is
that the interpolation of the neighbours (if any) would result in
a putative ray good enough for an effective inlier selection. In
Fig. 2 we show the bundles of projector rays obtained after
calibration. The first image depicts the bundle obtained by
calibrating the projector with a standard pinhole model. This
happens by using the estimated Ce as the 3D points of a virtual
calibration objects and the associated projector codes as their
reprojections. While this could seem a good approximation, we
will show in the experimental section that the pinhole model
is not able to fully deal with the imperfection of commercial
quality lenses (as also observed in [15]). The other two show
the bundles obtained using the described unconstrained model
respectively before and after outlier filtering and ray correction.
Fig. 2. The bundles of rays that can be obtained after calibration of the projector using the reconstructed 3D points. In the first image we adopted the
pinhole+distortion model. The second and third image show the results obtained using the unconstrained model respectively with and without outlier correction.
Note that the pinhole model is able to calibrate all the rays, while the unconstrained model can be populated only by the rays that hit the scanned surface, thus
they are a bit less. Also note that all the miscalibrated rays have (apparently) disappeared after outlier removal.
4082
Increment in coverage
1.7
1.6
1.5
1.4
1.3
Ganesha unconstrained
Ganesha pinhole
Sphere unconstrained
Sphere pinhole
1.2
1.1
Fig. 4. Scanner head with 2 calibrated cameras and an uncalibrated projector.
III.
•
E XPERIMENTAL E VALUATION
In order to evaluate the proposed method we built an
experimental setup similar to many off-the-shelf structured
light 3D scanners (see Fig. 4). We accurately calibrated the
two 1280x1024 pixels cameras for both intrinsic and extrinsic
parameters according to both the pinhole model and to the
unconstrained model. The projector used is an SVGA Dlp
micro projector. We implemented three reconstruction models:
•
Baseline: the unconstrained stereo camera reconstruction model that works without needing projector calibration presented in [15]. We expect this to be the
most accurate but to exhibit less coverage;
•
Pinhole: a reconstruction configuration that uses the
projector calibrated according to the pinhole model
(including distortion) to enhance coverage. We expect
this to be less accurate due to the limitation of the
pinhole model, especially for the cheap optics of
commercial-quality projectors;
Fig. 3.
Fig. 5.
5
10
15
20
25
Number of shots
30
35
40
Increment in the coverage with respect to the number of scans.
Unconstrained: the reconstruction model using the
unconstrained projector calibrated with the approach
described in this paper.
We tested these models by scanning three different objects:
a small figurine of Ganesha, which exhibits small details and
thus many high frequency areas, a regular sphere, which is
rather smooth and includes only low frequencies, and finally
a flat plane, used as a featureless reference object. For each
object we acquired about 100 scans covering almost all the
projector frustum, and we compared the results obtained by
calibrating the projector with different amounts of randomly
selected shots subsets. We adopted four different evaluation
criteria that are described in the following subsections.
A. Enhanced Coverage
With this test we measure the ratio between the area of
the surface obtained with the calibrated projector methods
and with baseline. This metric represents the enhancement in
Coverage difference between the baseline (top row) and the unconstrained method (bottom row) for some different subjects.
4083
2.4
Unconstrained
ned
Pinhole
2.2
2
RMS w.r.t. baseline
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
5
10
15
20
25
Number of shots
30
35
40
Fig. 6. Accuracy of the reconstruction with respect to the baseline method. The close-ups on the left part of the figure show a detail of the reconstruction
obtained respectively with the baseline, unconstrained and pinhole methods.
terms of coverage. In Fig. 5 we plot the coverage increment
with respect to the number of shots used to calibrate (which
correspond to the total number of scans, since the method is
online). Of course we show only the curves for the Sphere
and Ganesha objects, since there is no increment for the
plane (which is equally well seen by both cameras). Note that
the latter object obtains a larger advantage since it contains
many convoluted areas that are hard to capture at the same
time by two cameras and the projector. Note also that the
pinhole model reaches immediately the maximum increment
while the unconstrained model requires from 15 to 30 shots
to perform equally well. This is expected since in this case
the calibration includes all the rays from the start. However,
we will see in the next test that this advantage comes at the
cost of a lower accuracy. Some qualitative examples of the
coverage are shown in Fig. 3. Here the scattered edges of the
plane are due to the fact that not all the projector rays have
been recovered. This happens simply because the rays on the
periphery of the frustum appears in fewer scans of the subject,
which is expected. If a full coverage of the projection must be
guaranteed, this can be obtained offline using a bigger planar
object encompassing the whole frustum.
B. Reconstruction accuracy
To give a reasonable accuracy measure, we decided to
adopt the baseline method as a reference. This is a reasonable
choice since we already discussed in [15] the accuracy of
a camera pair calibrated with the unconstrained model. In
this regard, we express the accuracy as the RMS error, after
ICP registration [19], between the acquired surface and the
”ground truth” offered by the baseline. Note that such RMS is
expressed in world unit, which, since the cameras have been
calibrated with a computer monitor, corresponds to the size
of a pixel on that specific screen (approximately 0.2 mm). In
Fig. 6 we show the obtained accuracy after different amounts
of scans. The pinhole method requires few shots to reach its
maximum accuracy. However it always performs worse that
the unconstrained method. Furthermore the standard deviation
of the pinhole curve is narrower. These phenomena can be
explained respectively by the fact that the pinhole model is not
able to fully handle the imperfections of a real lens and that its
statistical nature makes it very stable with respect to the set of
shots selected for calibration. The unconstrained method, albeit
after several shots, allows for a significantly better accuracy.
0.55
Baseline
Unconstrained
Pinhole
0.5
0.45
ICP RMS error
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
5
10
15
20
25
Number of shots
30
35
40
Fig. 7. Repeatability of the reconstruction for different scans of the same subject. On the right part of the figure we show some slices from the acquired meshes
to illustrate the alignment between subsequent scans respectively with the baseline, pinhole and unconstrained methods.
4084
R EFERENCES
RMS error w.r.t. fitted plane
1.2
Baseline
Unconstrained
Pinhole
1
[1]
[2]
0.8
0.6
[3]
0.4
0.2
5
10
15
20
25
Number of shots
30
35
[4]
40
Fig. 8. Coherence of the reconstructed surface with a planar reference target.
[5]
C. Surface Repeatability
While the accuracy measures the compliance of the results
with respect to the ground truth, we are also interested in the
repeatability of the reconstruction within the same method.
To evaluate this metric we took several scans of the same
subject with slightly different poses and we computed the
average RMS error, after ICP registration, between the surfaced
acquired using the same method. Basically, this measure gives
us an insight about the resilience of the method to random
noise and to aliasing error generated by the interplay between
camera and projector rays. In Fig. 7 we plot such measure for
the baseline method (which appears as a horizontal line since
it does not depends on the number of scans) and of the other
two methods. We can conclude that all the tested approaches
exhibit a good repeatability, in the order of hundredths of
a millimetre. This repeatability appears to be not strongly
sensitive to the number of scan used to calibrate, with the
possible exception of the pinhole method that performs less
well with few shots.
[6]
[7]
[8]
[9]
[10]
[11]
[12]
D. Planarity
Finally, we measured the planarity of the reconstruction of
a reference plane made by coating with matte paint a float
glass. This is done by computing the average RMS error with
respect to a a general plane that has been fitted to the data.
The rationale of this measure is to assess spatial distortions
that usually characterizes imperfect calibrations. The results
obtained are shown in Fig. 8. We can observe that the pinhole
method produces the surface with larger distortion. This is
certainly attributable to the inherently imperfect correction of
the distortion.
IV.
C ONCLUSION
With this paper we introduced an online projector calibration method based on the unconstrained imaging model that
can be seamlessly applied to many commonly available 3D
scanning systems. The main advantage of this method is that it
can be performed during the normal scanning process, allowing
an improved scene coverage with little or no additional effort.
Furthermore, we have shown by extensive experiments that
an increase ranging from 30 to 60 percent in the recovered
surface area can be easily obtained without sacrificing the
reconstruction accuracy.
[13]
[14]
[15]
[16]
[17]
[18]
[19]
4085
C.-Y. Chen and H.-J. Chien, “An incremental target-adapted strategy
for active geometric calibration of projector-camera systems,” Sensors,
vol. 13, no. 2, pp. 2664–2681, 2013.
X. Chen, J. Xi, Y. Jin, and J. Sun, “Accurate calibration for a cameraprojector measurement system based on structured light projection,”
Optics and Lasers in Engineering, vol. 47, no. 34, pp. 310 – 319, 2009.
J. Huang, Z. Wang, J. Gao, and Q. Xue, “Projector calibration
with error surface compensation method in the structured
light three-dimensional measurement system,” Optical Engineering,
vol. 52, no. 4, pp. 043 602–043 602, 2013. [Online]. Available:
http://dx.doi.org/10.1117/1.OE.52.4.043602
D. Moreno and G. Taubin, “Simple, accurate, and robust projectorcamera calibration,” in 3D Imaging, Modeling, Processing, Visualization
and Transmission (3DIMPVT), 2012 Second International Conference
on, 2012, pp. 464–471.
S. Audet and M. Okutomi, “A user-friendly method to geometrically
calibrate projector-camera systems,” in Computer Vision and Pattern
Recognition Workshops, 2009. CVPR Workshops 2009. IEEE Computer
Society Conference on, 2009, pp. 47–54.
J. C. Lee, P. H. Dietz, D. Maynes-Aminzade, R. Raskar, and S. E.
Hudson, “Automatic projector calibration with embedded light sensors,”
in Proceedings of the 17th Annual ACM Symposium on User Interface
Software and Technology, ser. UIST ’04. New York, NY, USA: ACM,
2004, pp. 123–126.
M. Kimura, M. Mochimaru, and T. Kanade, “Projector calibration using
arbitrary planes and calibrated camera,” in Computer Vision and Pattern
Recognition, 2007. CVPR ’07. IEEE Conference on, 2007, pp. 1–8.
K. F. H. Yousef B. Mahdy and M. A. Abdel-Majid, “Projector calibration using passive stereo and triangulation,” International Journal
of Future Computer and Communication, vol. 2, no. 5, pp. 385 – 390,
2013.
T. Okatani and K. Deguchi, “Autocalibration of a projector-camera system,” IEEE Transactions on Pattern Analysis and Machine Intelligence,
vol. 27, no. 12, pp. 1845–1855, 2005.
R. Hartley, R. Gupta, and T. Chang, “Stereo from uncalibrated cameras,”
in Computer Vision and Pattern Recognition, 1992. Proceedings CVPR
’92., 1992 IEEE Computer Society Conference on, 1992, pp. 761–764.
O. Faugeras, Three-dimensional Computer Vision: A Geometric Viewpoint. Cambridge, MA, USA: MIT Press, 1993.
D. Fofi, J. Salvi, and E. Mouaddib, “Uncalibrated vision based on
structured light,” in Robotics and Automation, 2001. Proceedings 2001
ICRA. IEEE International Conference on, vol. 4, 2001, pp. 3548–3553.
R. Furukawa and H. Kawasaki, “Dense 3d reconstruction with an
uncalibrated stereo system using coded structured light,” in Computer
Vision and Pattern Recognition - Workshops, 2005. CVPR Workshops.
IEEE Computer Society Conference on, 2005, pp. 107–107.
S. Yamazaki, M. Mochimaru, and T. Kanade, “Simultaneous selfcalibration of a projector and a camera using structured light,” in
Computer Vision and Pattern Recognition Workshops (CVPRW), 2011
IEEE Computer Society Conference on, 2011, pp. 60–67.
F. Bergamasco, A. Albarelli, E. Rodol`a, and A. Torsello, “Can a
fully unconstrained imaging model be applied effectively to central
cameras?” in Proceedings of the 2013 IEEE Conference on Computer
Vision and Pattern Recognition, ser. CVPR ’13. Washington, DC,
USA: IEEE Computer Society, 2013, pp. 1391–1398. [Online].
Available: http://dx.doi.org/10.1109/CVPR.2013.183
E. Lilienblum and B. Michaelis, “Optical 3d surface reconstruction by
a multi-period phase shift method,” Journal of Computers, vol. 2, no. 2,
pp. 73–83, 2007.
K. Shoemake, “Animating rotation with quaternion curves,” in Proc.
of the 12th annual conference on Computer graphics and interactive
techniques, 1985, pp. 245–254.
ˇ ara, “Dual quaternions
L. Kavan, S. Collins, C. O’Sullivan, and J. Z´
for rigid transformation blending,” Trinity College Dublin, Tech. Rep.
TCD-CS-2006-46, 2006.
P. J. Besl and N. D. McKay, “A method for registration of 3-d shapes,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 14, no. 2, pp. 239–256,
1992.